Dynamic prognostication and treatment planning for hepatocellular carcinoma: A machine learning-enhanced survival study using multi-centric data

Abstract

<p>A reliable system for dynamic prognostication and management of hepatocellular carcinoma (HCC) is urgently needed but currently unavailable. In our previous work, we developed a machine learning algorithm termed "Survival Path" (raw-SP) to enhance prognostication with longitudinal survival data. However, the previous model was limited to intermediate stage HCC patients, and it faced the risk of overfitting due to path proliferation. In this study, we developed a novel framework incorporating nodal fusion techniques to mitigate the risk of overfitting, and expanded the model's applicability to all stages of HCC patients. A post-fusion survival map (fusion-SP) containing 14 different paths was built, which demonstrated superior or non-inferior accuracy in dynamic prognosis prediction for HCC patients compared with raw-SP, as well as traditional staging systems within the first 15 months since initial diagnosis in large-scale derivation, internal and external validation cohorts. Subgroup analysis showed the fusion-SP demonstrated superior performance in dynamic prognostication compared to other models among patients with BCLC stage C disease and initial tumor burden above up-to-seven criteria. Under the framework of fusion-SP, novel and distinct optimal combination treatment strategies for advanced-stage HCC patients at different key nodes were uncovered, where traditional staging frameworks fall short. The fusion-SP framework could serve as a robust tool for facilitating dynamic prognosis prediction and treatment planning for HCC. Moreover, our streamlined methodology holds the potential to be applied across various types of cancers.</p>